.avif){kind=small}
Validated AI, from any source, delivered safely into your hospital
A model proven elsewhere is not yet proven on your patients. ZeroOne checks data readiness, re-validates the model on your hospital's own de-identified data, deploys it inside the clinical workflow with human review, and keeps monitoring after go-live. The process is aligned with the NIST AI Risk Management Framework and ISO/IEC 42001.
A model proven elsewhere is not yet proven on your patients
Most clinical AI is validated on populations, workflows and data formats that differ from yours. Performance measured at a large academic centre does not transfer automatically to a community hospital in Ontario or a regional clinic in Germany. Mid-sized hospitals rarely have the data engineering capacity to re-check a model before it reaches clinicians. That is the gap this service closes.
Population shift
Age mix, comorbidities, language and referral patterns change what a model sees. We measure performance on your own de-identified records, not on the vendor's benchmark.
Workflow fit
An accurate alert that fires at the wrong moment, or too often, gets ignored. We test where the output lands, who acts on it and how long it takes.
Accountability
Every model gets a named clinical owner, a defined human review point and an escalation path before it is switched on.
Seven checks before a model reaches a clinician
Each check produces a number, a threshold agreed with your clinical lead, and a plain-language note on what it means for the service. Together they feed one report and one recommendation: go, conditional go, or no-go.
Sensitivity and specificity
How often the model catches true cases and how often it correctly leaves the rest alone, computed on your de-identified records against your documented outcomes.
A model tuned for a high-prevalence setting can miss cases or flood staff with false alarms in yours.
Positive predictive value
Of every alert the model raises, how many turn out to be real. We report it at the threshold the vendor recommends and at the threshold your team would actually use.
Positive predictive value drives alert fatigue. When most alerts are false, clinical teams stop reading them.
Calibration
Whether a predicted 20 percent risk really happens about 20 percent of the time in your population, checked across risk bands.
Poorly calibrated scores push clinicians to over-treat or under-treat even when the ranking of patients is good.
Subgroup performance
The same metrics broken down by age band, sex, site and other locally relevant groups, with sample sizes shown so that small groups are not over-read.
A model can look fine overall and still underperform for the patients who need it most. This is the bias check hospitals are increasingly asked to document.
Drift over time
Performance recomputed on recent data and compared with the validation baseline, with a defined trigger that sends the model back for review.
Patient mix, documentation habits and upstream systems change. A model that was safe last year may not be safe now.
Workflow fit
Where the output appears, who sees it, the time from alert to action, and how often clinicians override or dismiss it, measured in silent mode before go-live.
Accuracy without adoption delivers nothing; adoption without accuracy is a risk.
Governance readiness
A named clinical owner, an approved use statement, a human review point, an escalation path and a retirement rule, all written down before activation.
Regulators and accreditation bodies increasingly expect this file to exist. It is also what makes a no-go decision possible later.
Four stages from data readiness to continuous monitoring
Data readiness
We assess whether the data the model needs exists, is complete enough and can be accessed lawfully, then document the gaps. Output: a data readiness register.
Local validation
We run the model on your de-identified data, compute the seven checks and report by subgroup. Output: a validation report with a go, conditional-go or no-go recommendation.
Workflow deployment
We connect through approved interfaces, define the human review point and run a silent-mode period before clinicians see the output. Output: a deployment and governance pack.
Continuous monitoring
We track drift, alert volume, override rate and incidents against agreed thresholds and report monthly to the named owner. Output: a monitoring dashboard and review log.
Validation supports, but does not replace, a manufacturer's regulatory obligations. Where a solution is regulated as software as a medical device (SaMD), our report is evidence for your local governance, not a certification.
Aligned with recognised frameworks, without claiming certification
Our assurance process follows the four functions of the NIST AI Risk Management Framework (Govern, Map, Measure, Manage) and the management-system approach of ISO/IEC 42001. Where analytics maturity is in scope, findings are mapped to the HIMSS Analytics Maturity Assessment Model (AMAM). For hospitals in the European Union we also track the obligations the EU AI Act places on high-risk clinical systems. We describe our work as aligned with these frameworks; we do not claim certification against them.
Three deliverables, written for clinical and executive readers
Validation Report
Performance on your population for each of the seven checks, by subgroup, with the agreed thresholds and a go, conditional-go or no-go recommendation in plain language.
Model Card and Governance Pack
A one-page model card (purpose, data, limits, owner) plus the approved use statement, human review point, escalation path and retirement rule required before activation.
Monitoring Plan and Review Log
Thresholds for drift, alert volume and override rate, a monthly review template, and the trigger that returns a model to validation or retires it.
Sample documents are illustrative and use synthetic data. Real reports are produced for one hospital, one model and one agreed set of thresholds, and are shared only with that hospital.
What a validation report looks like
The example below is built from synthetic data for an illustrative inpatient deterioration-risk alert. It shows the shape of the report: overall performance, the same metrics by subgroup with sample sizes, and a decision per row against thresholds agreed with the clinical lead. Real reports use your hospital's de-identified data and your thresholds.
| Subgroup | n | Sensitivity | Specificity | PPV | Decision |
|---|---|---|---|---|---|
| All patients | 12,400 | 0.84 | 0.91 | 0.38 | Meets target |
| Age 18–64 | 7,900 | 0.86 | 0.92 | 0.40 | Meets target |
| Age 65+ | 4,500 | 0.79 | 0.89 | 0.35 | Below target: recalibrate |
| Female | 6,300 | 0.85 | 0.91 | 0.38 | Meets target |
| Male | 6,100 | 0.83 | 0.90 | 0.37 | Meets target |
| Community site | 3,200 | 0.77 | 0.90 | 0.33 | Below target: check data completeness |
Illustrative example with synthetic data. Model: inpatient deterioration-risk alert. Agreed thresholds: sensitivity ≥ 0.80, specificity ≥ 0.88, PPV ≥ 0.35. Recommendation: conditional go; run in silent mode for 8 weeks and recalibrate for patients aged 65 and over before activation.
.png)
.png){kind=small}
